The Quality Challenge in Watchmaking
A Geneva watchmaking subcontractor specializing in high-precision component machining for prestigious manufacturers produced thousands of tiny parts daily (shafts, bridges, gear wheels). Quality was absolutely critical as the slightest defect on a part could compromise the functioning of a watch worth tens of thousands of francs. Quality control was performed by manual visual inspection under microscope—a slow process, tiring for inspectors, and imperfect.
Despite experienced inspectors, about 0.5% of defective parts went unnoticed and were delivered to clients, generating costly complaints and threatening business relationships. Manual inspection also slowed production as each batch had to wait to be inspected before shipping. The company sought a solution enabling automatic, exhaustive, real-time control without slowing production pace.
The Computer Vision Solution
We developed a complete quality control system using computer vision combining Azure Custom Vision, Azure IoT Edge, and Power BI. The architecture integrates directly into the existing production chain.
Each production station was equipped with a high-resolution image capture system comprising a 12-megapixel industrial camera with macro lens, multi-directional LED lighting eliminating shadows and reflections, and a motorized turntable enabling capture of the part from 8 different angles. Images are captured automatically as soon as a part is placed on the tray, without human intervention or production slowdown.
Images are analyzed in real-time by a computer vision model trained specifically to detect defect types encountered in this production: scratches on polished surfaces, burrs on edges, threading defects, dimensions out of tolerance, abnormal curvatures, and contaminations (dust, machining residue). The model was developed with Azure Custom Vision and trained on over 50,000 images of conforming parts and 8,000 images of defective parts of all types.
To handle latency constraints (decision in under 2 seconds), the model is deployed in edge computing on an Azure IoT Edge device installed directly in the workshop. This architecture avoids sending images to the cloud, guaranteeing instant response times even in case of network problems. The model runs on a processor with GPU acceleration enabling rapid analysis of 8 images per part.
When the system detects a defect, several automatic actions trigger. The defective part is automatically ejected from the production flow to a reject bin, a red light turns on to alert the operator, a photo of the defect is saved in Azure Blob Storage with metadata (defect type, station, timestamp, batch number), and a notification is sent to the quality manager via Teams if the defect rate exceeds a critical threshold suggesting a machine problem.
Conforming parts automatically continue their path in the production process. A unique QR code is laser-engraved on each approved part, enabling complete traceability. This code links the physical part to its digital file containing inspection photos, production parameters, and conformity certificate.
A Power BI dashboard connected in real-time to IoT Edge devices displays quality metrics for each production station: number of inspected parts, conformity rate, detected defect types, and hourly trends. This dashboard enables quick identification of quality drifts and intervention before a large number of defective parts is produced.
Measured Results
After ten months of progressive deployment across all production lines, results are exceptional. The rate of defects shipped to clients dropped 98%, from 0.5% to 0.01%. Client complaints practically disappeared, considerably strengthening the subcontractor's reliability reputation. Several prestigious manufacturers increased their orders following this qualitative improvement.
Productivity increased 35%. Automatic in-line inspection eliminates bottlenecks from deferred manual control. Parts are validated immediately and can be shipped without waiting. The cycle time from quote to delivery decreased by 2 days on average.
Non-quality costs fell 85%. Defective parts are detected immediately after production rather than at process end or worse, at the client's. This avoids rework costs, return logistics, and complaint processing. Annual savings exceed 180,000 CHF.
Quality control teams were reassigned to higher value tasks: root cause analysis of defects, continuous process improvement, and new part validation. Their professional satisfaction improved because they no longer spend their days looking through a microscope—repetitive and tiring work.
Complete traceability enabled by QR codes and image archiving has multiple benefits. In case of later problems, the subcontractor can prove the part was conforming at shipment. This capability already helped avoid an unjustified complaint, saving a business relationship worth several hundred thousand francs annually.
Continuous Model Improvement
The computer vision model is retrained quarterly with newly collected images, including rare false positives (conforming parts rejected by error) and false negatives (missed defects) reported. Initial accuracy of 94% progressed to 99.2% after ten months of continuous learning.
When a new part type enters production, a few-hour calibration phase adapts the model to the geometric and aesthetic specificities of this new reference. The system now handles over 150 different references with a single installation.
A random human check mechanism verifies 1% of accepted parts to detect any false negatives. These rare cases are immediately used to retrain the model, creating a permanent improvement loop.
Architecture and Industrial Integration
The architecture combines SICK industrial cameras with Chromasens lighting, Azure IoT Edge devices based on NVIDIA Jetson for edge compute with GPU acceleration, Azure IoT Hub for telemetry and device management, Azure Blob Storage for defect image archiving, and Power BI for real-time dashboards.
Integration with existing production machines was achieved via Siemens programmable controllers communicating in Profinet. The vision system inserts naturally into the flow without major modification of existing equipment.
Reliability is critical in a production environment. The edge architecture guarantees production can continue even in case of network or cloud failure. IoT Edge devices function autonomously and synchronize data when connectivity is restored.
Total system cost including vision equipment (8 stations), IoT Edge devices, Azure licenses, and houle support represents an initial investment of 120,000 CHF and a monthly operating cost of 800 CHF. ROI was achieved in 14 months through non-quality cost reduction and production volume increase.
Extension to Other Controls
Building on visual control success, the company now deploys similar systems for other quality aspects: automatic dimensional control by calibrated vision replacing manual calipers, detection of metallurgical defects by thermographic imaging, and inspection of polished surfaces by grazing light to detect micro-scratches invisible to the naked eye.
Conclusion
This computer vision quality control system demonstrates how AI can surpass human capabilities for repetitive inspection tasks requiring precision and exhaustiveness. By guaranteeing perfect quality while accelerating production, the watchmaking subcontractor transformed quality into a differentiating competitive advantage. This excellence became a major commercial argument in a very demanding market.